An Integrated Statistical Framework for Lesion Detection Using Dynamic PET

点击翻译按钮获取中文摘要

DESCRIPTION (provided by applicant): Positron emission tomography (PET) with FDG has become a widely accepted and used clinical molecular imaging tool for disease diagnosis, staging, treatment planning, management and evaluation. Although conventional static PET imaging provides high sensitivity in tumor detection, further improvement is important since even a small percentage of false negatives can have a major impact on treatment, cost and outcome. Visual inspection of static images is potentially inaccurate for small tumors due to limited spatial resolution and low lesion-to-background contrast. Computer aided detection (CAD) combined with use of dynamic PET data could assist in improving sensitivity and specificity for these small lesions. The goal of this exploratory Bioengineering Research Grant proposal is to investigate such a CAD method for dynamic FDG PET that integrates image reconstruction, lesion detection and thresholding in a statistical framework. The method will be optimized based on the properties of the dynamic PET data and the imaging system, and is designed to use standard dynamic data sets without the need for a measured blood input function. The CAD system will automatically provide a voxel-wise statistical map indicating probable lesion locations. By using a statistical detection algorithm that combines spatial and temporal information, we expect to be able to improve detection of small lesions that are not clearly visible in standard static scans and thereby provide improved diagnostic information to the radiologist. We will apply our maximum a posteriori (MAP) approach to PET image reconstruction to data from the new generation of clinical scanners, and optimize performance in terms of modeling and calibration procedures based on the characteristics of the scanner. The resulting images of estimated dynamic tracer uptake, as well as their approximate covariance, computed based on a theoretical analysis of the reconstruction algorithm, will be used as input to a matched subspace detector. This detector characterizes typical tumor and normal tissue dynamics using linear subspaces in combination with a generalized likelihood ratio test, to generate a voxel-wise statistical map indicating the likelihood of tumor presence or absence. Typical tumor and normal tissue subspaces will be obtained using a training dataset from multiple subjects with tumor and normal tissue regions of interest (ROIs) identified by a radiologist. The statistical detection map will then be thresholded to obtain a voxel-wise indication of likely tumor locations, while controlling for the effects of multiple comparisons. We will implement, optimize and perform preliminary evaluation of this CAD approach for dynamic data collected at USC using the Siemens Biograph TruePoint scanner. Evaluation will use Monte Carlo simulation and retrospective human studies. Human studies will focus on patients with liver metastases from colorectal cancer who are enrolled in an ongoing clinical trial. Serial imaging studies, with subsequent surgical resection and independent verification through pathology and intraoperative ultrasound, will provide a basis to evaluate the performance of our CAD detection approach. PUBLIC HEALTH RELEVANCE: Positron Emission Tomography (PET) has been widely used in cancer diagnosis, staging, treatment planning, management and evaluation. One of the main functions of PET is to detect tumors and metastatic lesions, which is conventionally done by visual inspection of a static volumetric image by a radiologist. This project is focused on using multiple images of the patient collected in a single session, in combination with a novel computer aided detection (CAD) method, to assist radiologists in detecting small tumors that may not be clearly visible using standard imaging protocols. Success of this project may lead to improved detection, staging and monitoring of metastatic disease.